P1A-5 Analysis of Backscattered Signals with a Neural Network Model for Microemboli Classification

Emboli classification is of high clinical importance for selecting appropriate patient treatment. Several ultrasonic (US) methods using Doppler processing have been used for emboli detection and classification as solid or gaseous matter. We suggest in this experimental study to exploit the radio-frequency (RF) signal backscattered by the emboli since it contains additional information about the embolus than the Doppler signal. The aim of the study is the analysis of RF signals using radial-basis function neural network (RBFN) in order to classify emboli. Anthares scanner with RF access was used with a transmit frequency of 1.82 MHz and mechanical indices (MI) varying from 0.2 to 1.1. To imitate emboli US behaviour, Sonovue microbubbles were injected in a nonrecirculating flow phantom with a 0.8 mm diameter vessel with a constant flow. Non perfused tissue was assumed to behave as solid emboli. Sonovue concentration was chosen such that fundamental scattering from tissue and contrast were identical. The training, validation and test sets consisted of 50, 26 and 26 samples respectively for each MI. The training set was used to determine the network parameters while the optimal number of centres in the hidden layer of our RBFN was determined from the validation set. Amplitudes and bandwidths of fundamental, 2nd and 3rd harmonic components as well as the parameters of Gaussian approximation of the frequency spectrum were the selected input parameters of the RBFN model. When the bandwidths or the amplitudes of the fundamental and the harmonic frequencies were used as input parameters, the neural network led to unsatisfactory classification rates: only 31% of gaseous emboli were correctly characterized at low MI. This rate was increasing with the MI up to 100% of correct classification. Gaseous and solid emboli were 100% correctly characterized when the Gaussian coefficients were selected as input parameters, independently of the MI. The study has demonstrated the opportun- ity to classify emboli based on a RF signals neural network analysis.